Is Turnitin's AI detector accurate?
Turnitin led the early independent tests at 76% and reports a document false positive rate under 1%, but simple evasion edits cut its accuracy from 50% to 7.9% in a 2024 study.
Short answer
Turnitin's AI detector did best in early independent tests, but it is not reliable enough to prove misconduct. It ranked first of 14 tools at 76% in 2023, yet a 2024 study found simple evasion edits cut its accuracy from 50% to 7.9%. Turnitin puts its own document false positive rate under 1%.
Is Turnitin's AI detector accurate? It is the detector most students meet, because their school runs it on every submission, so the answer matters more here than for any other tool. The published record has two sides. On unedited AI text, Turnitin led the field in the academic tests that included it. On text that has been reworded, even lightly, its score collapsed. Its false positive figures come from Turnitin itself. Every number below is dated and linked to its source, and vendor figures are labelled. For how we weigh sources, see how we score AI accuracy.
Turnitin AI detection test results
- Independent study
- Vendor claim
| Test | What was measured | Result | Date | Source |
|---|---|---|---|---|
| Turnitin internal testingVendor's own figure, not independently checked | False positive rate, documents with 20% or more AI writing | Less than 1% | Source 1Vendor claim | |
| Turnitin internal testingVendor's own figure, not independently checked | False positive rate, sentence level | About 4% | Source 1Vendor claim | |
| Weber-Wulff et al. 2023, 54 documents | Detection rate (binary accuracy) | 76%, 1st of 14 tools | Source 3Independent study | |
| Perkins et al. 2024 | Detection rate on unedited AI text | 61%, 2nd of 7 detectors | Source 4Independent study | |
| Perkins et al. 2024 | Detection rate after evasion edits (average of three scoring approaches) | 7.9%, down from 50% | Source 4Independent study |
Turnitin has also published how often its detector fires. Of more than 200 million papers reviewed in its first year, over 22 million, about 11%, had at least 20% AI writing, and about 3% had 80% or more, as of 2024-03-21. That is a usage figure from Turnitin, not a measure of accuracy.
What the Turnitin AI detection accuracy numbers mean
Start with the good news for Turnitin. Weber-Wulff and colleagues ran 54 documents through 14 detection tools in 2023, and Turnitin came top at 76%. The authors still noted it stayed below 80% and struggled with paraphrased and edited AI text. In 2024, Perkins and colleagues found it caught 61% of unedited AI text, second of seven detectors behind Copyleaks at 64.8%.
Then the bad news. In the same Perkins study, the researchers applied simple adversarial rewrites to AI text. Turnitin's accuracy fell from 50% to 7.9%, a drop of 42.1 points and the largest of the seven detectors. In plain terms, light editing defeated Turnitin most of the time in that test.
Turnitin's false positive rate comes from its own testing. The company says fewer than 1% of documents with 20% or more AI writing are wrongly flagged at the document level, and that the sentence-level false positive rate is about 4%. It also says 54% of false positive sentences sat next to real AI writing. That matters for students: a highlighted sentence can be human writing that happens to sit beside AI text.
Both academic studies tested older versions of Turnitin's detector, and both papers warn their results may not reflect current tools.
Can Turnitin detect ChatGPT?
Often, when the AI text is pasted in unchanged. Turnitin caught 61% of unedited AI text in Perkins et al. and led the 2023 Weber-Wulff test at 76%. Once the same kind of text is reworded, the picture changes: its accuracy fell to 7.9% after evasion edits. The studies measured AI-generated text, which is what a pasted ChatGPT answer is: Turnitin catches much of it unedited and misses most of it once reworded. Neither result tells a teacher whether a particular student used ChatGPT.
Where Turnitin goes wrong
Edited AI text
Its accuracy dropped from 50% to 7.9% after evasion edits, the steepest fall of seven detectors.
Sentence-level flags
By Turnitin's own figures, about 4% of human sentences get flagged, and 54% of those sit next to AI writing.
Mixed documents
Its under 1% document figure applies to documents with 20% or more AI writing, so the threshold shapes what the number means.
Human control texts
Across all seven detectors in Perkins et al., human control samples were classified correctly only 67% of the time.
Scale
With over 200 million papers reviewed in a year, by Turnitin's own count, even a small false positive rate touches many students.
How Turnitin compares in the same tests
Only compare detectors inside one study.
| Test | Turnitin | Others in the same test |
|---|---|---|
| Weber-Wulff et al. 2023, binary accuracy | 76% (1st of 14) | ZeroGPT 59%, GPTZero 54% |
| Perkins et al. 2024, unedited AI text | 61% (2nd of 7) | Copyleaks 64.8%, ZeroGPT 46.1%, GPTZero 26.3% |
| Perkins et al. 2024, after evasion edits | 7.9% (from 50%) | ZeroGPT 17.3% (from 31.3%); all seven averaged 22.2% |
Turnitin beat ZeroGPT and an older GPTZero on unedited text in both academic tests, then fell below ZeroGPT once the text was edited. Newer results for GPTZero and Grammarly come from the RAID leaderboard, a different test, so they cannot be ranked against these figures. For the full picture, see are AI detectors accurate.
Non-native English writers and AI detectors
The best known fairness study, by Liang and colleagues at Stanford, found seven detectors flagged non-native English essays as AI 61.22% of the time on average (Liang et al., 2023), with 97.8% of 91 TOEFL essays flagged by at least one detector. Turnitin was not one of the seven. The authors link the bias to simpler vocabulary, which detectors read as predictable, and that is how detectors in general score text. If English is your second language and Turnitin flags your work, this study is worth raising.
What a Turnitin AI score can and cannot prove
No detector score should be the only evidence in a misconduct case.
It can show
- A Turnitin AI score can tell you that parts of a submission read like model output to Turnitin's detector. That can be a reason to read the highlighted sentences closely and ask the writer about them.
It cannot prove
- A Turnitin AI score cannot tell you who wrote the work, how AI was used, or whether a flagged sentence is human writing that happens to sit beside AI text. With a sentence-level false positive rate of about 4% by Turnitin's own figure, a highlight on its own proves little.
What to do if Turnitin flags your work
Ask your instructor for the AI writing report, including the percentage and the highlighted sentences.
Collect your drafts, outline, notes, sources and version history from Google Docs or Word.
Point to Turnitin's own sentence-level false positive rate of about 4%.
If English is not your first language, cite the Liang et al. findings on non-native writers.
Offer to walk through your argument and sources, or to write a short sample under supervision.
Ask your school for its appeal process in writing, and ask that any decision rest on more than one piece of evidence.
Another detector to try
Another option is the free AI detector from Walter Writes. Its page says checks are free and unlimited, that it detects text from ChatGPT and Gemini, that it gives a percentage score for every scan and that it works in more than 20 languages. You can try Walter's AI detector here. A second detector's score is one more signal, not proof either way.
Sources
- Vendor claim
- Vendor claim
- Independent study
- Independent study
- Independent study
Questions
How accurate is Turnitin's AI detector?
It depends on the text. Turnitin led the 2023 Weber-Wulff test of 14 tools at 76% and caught 61% of unedited AI text in a 2024 study. After simple evasion edits, its accuracy fell from 50% to 7.9%. Both studies used older versions of the detector.
What is Turnitin's AI false positive rate?
By Turnitin's own figures, under 1% at the document level for documents with 20% or more AI writing, and about 4% at the sentence level. Turnitin also says 54% of false positive sentences sat next to real AI writing. These are vendor numbers from 2023, from Turnitin's internal testing.
Can Turnitin detect ChatGPT if I paraphrase it?
Much less often. In Perkins et al. (2024), simple adversarial rewrites cut Turnitin's accuracy from 50% to 7.9%, the biggest drop of the seven detectors tested. Unedited AI text is where Turnitin performs best; reworded text is where it performs worst. Either way, a score alone does not show who wrote the work.
Can Turnitin flag human writing as AI?
Yes. Turnitin's own sentence-level false positive rate is about 4%, and its document-level rate is under 1% for documents above its 20% threshold. Across seven detectors in a 2024 study, human control samples were classified correctly only 67% of the time. Keep your drafts so you can show your process.
Should a school rely on a Turnitin AI score?
Not on its own. Turnitin led the early independent tests, but light editing defeated it in 2024, and its false positive figures come from its own testing. Over 200 million papers went through the detector in its first year, so even a small error rate affects many students. Use the score to start a conversation.
